OS-DS tracker: Orientation-variant Siamese 3D tracking with Detection based Sampling

Abstract

LIDAR point clouds are less affected by the weather and possess more depth of field than 2D images. However, sparsity and disorder of point clouds bring challenges for 3D object tracking due to the difficulty in detecting and encoding. In addition, ineffective candidate proposals increase the phenomenon of the loss of target tracking or wrong target tracking. In this paper, we propose an Orientation-variant Siamese 3D object tracking network that utilizes a Detection based Sampling module to generate candidate proposals (OS-DS tracker). The Detection based Sampling module is used to cut the excessive and useless proposals. And the multivariate Gaussian sampling generates the candidate proposals when the objects are not detected. Concretely, we first pre-train PointRCNN Network to globally detect objects from LIDAR point clouds. Then the 3D detected objects are refined by Candidate Regions Sampling to generate candidate proposals. Meanwhile, to make the feature vectors more discriminative, we design an Orientation-variant Siamese Auto-encoder, i.e., tracking loss regress to the intersection over union (IOU) of the template box and the candidate region boxes. Our method is tested on the KITTI dataset and the SHIP Tracking dataset. Compared with the previous state-of-the-arts, the proposed method outperforms in vessel tracking with 5.2%/2.1% improvement in Success and Precision.